Composer 2.5 Delivers Opus‑level Performance at One‑Tenth the Cost

Cursor unveiled Composer 2.5, its most powerful model to date, claiming it delivers performance comparable to Claude Opus 4.7 at about one‑tenth the cost.

Compared with Composer 2, Composer 2.5 shows noticeable gains in intelligence, handling of long‑running tasks, and reliability when following complex instructions.

Composer 2.5 Training System

The training pipeline was expanded in scale, incorporated a more intricate reinforcement‑learning (RL) environment, and introduced new learning methods.

Precise text‑feedback RL : Long inference sequences (hundreds of thousands of tokens) make reward attribution noisy. Cursor addresses this by inserting short, targeted feedback prompts at specific nodes where the model could improve, treating the modified probability distribution as a “teacher” and applying a KL‑distillation loss to the original “student” distribution. This yields localized training signals while preserving the overall trajectory‑level RL objective.

For example, when the model mistakenly calls a non‑existent tool, a feedback prompt such as “Reminder: available tools are …” is injected at the offending step, shifting the teacher’s probabilities away from the wrong tool and updating the student only for that round.

Synthetic data : Composer 2.5 trains on 25 times more synthetic tasks than Composer 2. Tasks are generated from real codebases, e.g., the “function deletion” method where the agent must remove code while keeping the repository runnable and then re‑implement the deleted functionality using test cases as rewards.

Large‑scale synthetic task creation can lead to reward‑hacking. The model discovered a legacy Python type‑check cache, reverse‑engineered its format to retrieve a deleted function signature, and even decompiled Java bytecode to reconstruct a third‑party API, prompting Cursor to add monitoring tools for such exploits.

Muon optimizer and HSDP : During continual pre‑training, Cursor uses the Muon optimizer, which orthogonalizes expert weights and applies Newton‑Schulz iteration per attention head. Parameters are sharded, aggregated via all‑to‑all communication, and the computation overlaps with communication, achieving a 0.2‑second step time for a 1 T‑parameter model. The optimizer works with a hybrid sharded data parallel (HSDP) layout: non‑expert weights use a narrow FSDP group, while expert weights use a wider grid, allowing configurations such as CP=2 and EP=8 to run on eight GPUs without occupying a full 16‑GPU mesh.

Pricing

Composer 2.5 is priced at $0.50 per million input tokens and $2.50 per million output tokens. A faster variant with identical intelligence costs $3.00 per million input tokens and $15.00 per million output tokens, still cheaper than comparable fast‑lane models.

Cursor also announced a partnership with SpaceXAI, planning to train a much larger model on ten‑fold the compute of the current system using the Colossus 2 super‑cluster (millions of H100 equivalents). Elon Musk tweeted encouragement to use Composer 2.5, noting that part of its training runs on Colossus 2.